ep 03 field notes
Show Us Your Agent Skills / EP 03 / guest dossier
ELEANOR BERGER JIMINY HEALTH HERMES · FNORD 157 SKILLS YOLO, FENCED

ELEANOR BERGER

Eleanor will not let one AI validate another's work: "there's this infinite regress where you could get stuck in a local minimum and you never get out of it." Which makes it a surprise that she now lets a personal agent roam free on a spare Mac mini named Fnord, publishing pages and running jobs it schedules itself. The same engineer who demands concrete verification is the one who fenced that machine off from everything that matters before giving it room to run.

EP 03 · ELEANOR BERGER · Fnord roaming free on a fenced Mac mini, live on stream

"If I know a little bit of what I'm doing, enough to evaluate it, all of a sudden I've got this exoskeleton."

Agents changed which projects are worth starting at all. "If something can't happen fast, I might just not do it." 00:38:03

FNORD INTRODUCED ITSELF

her personal agent harness: one Discord message in, a finished web page out

Minutes before the stream, Eleanor sent Fnord a single Discord message: "I want to present this Hermes instance to the audience. Create a cool interactive web page." That was the whole brief.

What came back, published to here.now on its own: a running catalogue of Fnord's skills, its parallel child agents, the GPT 5.5 route, and the Discord and WhatsApp channels it answers on.

The Fnord presentation page on here.now, listing its installed skills, child agents, GPT 5.5 model route, and Discord and WhatsApp channels
Fnord introduced itself as "a useful creature with a filesystem." The page it built to say so, published to here.now. [00:47:00]

"If you are getting instructions from the internet and you have connections out to the internet, you should assume that you're completely exposed."

Simon Willison's lethal trifecta, private data plus untrusted input plus an outbound network, taken literally. It is why Fnord lives on its own hardware and never gets near her clinical data at Jiminy. 00:56:57

GIVING IT MORE ROPE

her rules for letting an agent roam without getting burned
Earn autonomy graduallyStart at ask-for-everything, then hand over more rope as its self-evaluation proves reliable. "I gave it a bit more rope." Reversible the moment something crosses the line. 00:56:20
Segregate by hardware, not permissionsThe lethal trifecta is essentially unmitigable in software, so the fix is physical: a separate Mac mini over Tailscale, no path to clinical data. 00:42:51
Verify by realizing, not reviewingFor cloud infra, don't ask another model to bless the YAML. "Realize this configuration and sample what it actually created. That's something I can trust." 00:34:52
Under-specify the cron, let Hermes route"I create a lot of cron jobs and a lot of them I under-specify." Hermes decides when a script is enough, when to call an LLM inside it, and when the whole task needs a model. 01:00:51
Make publishing the default finish"Do a web page, publish it to here now." Sometimes she does not even ask and it publishes anyway, because her config says that is where results go. Dozens of pages a day. 00:49:17
Give scope a numberAgents read intent well and scope badly: ask for a one-pager, get a novel. "This needs to be 500 tokens or less. You need to give something concrete, otherwise it will go on and on." 00:39:55
Fnord's headline stats panel: 157 installed procedural skills, 13 active scheduled jobs, 123 max tool-calling turns, 3 parallel child agents, on the OpenAI Codex gpt-5-5 route
The rope, quantified: 123 max tool-calling turns, 3 parallel child agents, 13 scheduled jobs. Fnord reported its own bounds from scrubbed metadata, secrets omitted. [00:46:38]

"I asked very briefly in the chat and it invented this way to do it."

A one-line phone request on public transport, and Fnord built the whole Watch Later skill itself. It even wrote her in the third person, "I would never refer to myself in the third person," and she kept it anyway. 00:54:52
The youtube-watch-later-cache skill markdown in the Hermes skills browser, showing the cache root, key files, and run, refresh, and gist-catchup commands
Inside the skill Fnord generated: a cache, a transcript store, summary files, and run, refresh, and catch-up commands. She designed none of it. [00:54:55]

WHAT SHE LET LOOSE

the skills and the harness from her segment. install them, then make them yours: they're living documents
skill

here-now

Publishes HTML pages, files, and whole sites to live URLs without leaving the terminal. A first-party skill from here.now itself: "like gists, but for HTML."

skill

impeccable

Hands a coding agent a full frontend design language so it builds production-grade interfaces instead of generic ones. The design skill she leans on: "I know absolutely nothing about design."

skill

anki-connect

Drives Anki through the AnkiConnect API, gating every note- or card-modifying operation behind explicit confirmation. Runs as a background cron over her flashcard collection.

skill

youtube-watch-later

Reads your Watch Later queue from a logged-in browser, summarizes each video from its transcript, and caches what it has already processed so it never repeats work.

workflow

personal-agent-harness

The Fnord setup itself: a personal agent on spare hardware, reachable over Discord or WhatsApp, with autonomy granted gradually and a skill set the agent extends for itself.